Microsoft is in talks with Taiwan Semiconductor Manufacturing Co. (TSMC) to secure production capacity for more than 300,000 Maia 300 chips for delivery in 2027, according to Reuters, citing a report from The Information.
The next-generation accelerator may arrive as soon as September. Microsoft is also targeting capacity for more than 1 million Maia 300 chips over time, although component availability and negotiations with TSMC could make that goal difficult to reach.
Andrew Wall, general manager for Microsoft's Azure Maia, confirmed the company's continued investment in custom silicon but pushed back on the reported production figures.
"Microsoft continues to invest in custom silicon as part of our long-term AI infrastructure strategy. While we don't share production volumes, the figures reported don't reflect the scale of our program," Wall said, according to Reuters.
The company introduced its first Maia chip in November 2023 and launched the second-generation Maia 200 in January. Maia 200 was built by TSMC using its 3-nanometer process and includes a large amount of SRAM designed to help AI systems handle high volumes of requests.
Microsoft is chasing a crowded field
The Maia effort puts Microsoft in the same race as Google and Amazon, both of which started developing custom AI accelerators earlier and have moved further toward large-scale deployment.
Google has begun generating revenue from direct sales of its Tensor Processing Units, while Amazon's Trainium chips are gaining adoption. Anthropic, for example, has agreements involving both Amazon's Trainium and Google's TPUs, highlighting the scale Microsoft needs to reach if it wants its silicon to become a meaningful alternative.
Microsoft has also been discussing the possibility of supplying Maia chips to Anthropic, although no deal had been finalized, according to the reports. The broader goal is to reduce dependence on Nvidia's expensive and difficult-to-source GPUs while giving Microsoft greater control over the hardware powering its AI services.
The real test will be scale
Microsoft's Maia strategy is less about announcing another chip and more about proving it can manufacture and deploy custom silicon at hyperscale.
That makes the reported 300,000-chip target significant. If Microsoft can secure the manufacturing capacity and convince major cloud customers to use Maia, it could eventually have more control over AI computing costs and infrastructure.
But reaching more than 1 million chips would require reliable supplies of components and enough manufacturing capacity at TSMC.
There is also a competitive risk. Google and Amazon already have substantial experience deploying their own accelerators, while Nvidia remains the dominant supplier of AI computing hardware.
For Microsoft, Maia 300 therefore represents a bet on execution as much as chip design. A successful rollout could give Azure another tool to control the economics of AI workloads; delays or limited customer adoption could leave Microsoft still heavily dependent on Nvidia.
Also read: Taiwan prosecutors recently detained a reported Nvidia employee in an investigation into alleged smuggling of restricted AI servers into China.